Re: [PATCH] lib/stackdepot: Use page allocator if both slab and memblock is unavailable

[Marco Elver <elver@google.com>]

On Mon, 28 Feb 2022 at 11:05, Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
[...]
> > This is odd - who is calling stack_depot_init() while neither slab nor
> > memblock are available?
>
> It's not merged yet - but Oliver's patch (2/5) in his series [1] does:
> If user is debugging cache, it calls stack_depot_init() when creating
> cache.
>
> > @@ -4221,6 +4220,9 @@ static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
> >       s->remote_node_defrag_ratio = 1000;
> >  #endif
> >
> > +     if (s->flags & SLAB_STORE_USER && IS_ENABLED(CONFIG_STACKDEPOT))
> > +             stack_depot_init();
> > +
>
> Oliver's patch series enables stack depot when arch supports stacktrace,
> to store slab objects' stack traces. (as slub debugging feature.)
>
> Because slub debugging is turned on by default, the commit 2dba5eb1c73b
> ("lib/stackdepot: allow optional init and stack_table allocation by
> kvmalloc()") made stack_depot_init() can be called later.
>
> With Oliver's patch applied, stack_depot_init() can be called in
> contexts below:
>
>   1) only memblock available (for kasan)
>   2) only buddy available, vmalloc/memblock unavailable (for boot caches)
>   3) buddy/slab available, vmalloc/memblock unavailable (vmap_area cache)
>   4) buddy/slab/vmalloc available, memblock unavailable (other caches)
>
> SLUB supports enabling debugging for specific cache by passing
> slub_debug boot parameter. As slab caches can be created in
> various context, stack_depot_init() should consider all contexts above.
>
> Writing this, I realized my patch does not handle case 3).. I'll send v3.
>
> [1] https://lore.kernel.org/linux-mm/YhoakP7Kih%2FYUgiN@ip-172-31-19-208.ap-northeast-1.compute.internal/T/#t
> [2] https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/linux.git/log/?h=slub-stackdepot-v1
>
> > Do you have a stacktrace?
>
> Yeah, here:
>
> You can reproduce this on vbabka's slab-stackdepot-v1 branch [2] with
> slub_debug=U, and CONFIG_STACKDEPOT_ALWAYS_INIT=n
>
[...]
> [    0.000000] Call trace:
> [    0.000000]  __memset+0x16c/0x188
> [    0.000000]  stack_depot_init+0xc8/0x100
> [    0.000000]  __kmem_cache_create+0x454/0x570
> [    0.000000]  create_boot_cache+0xa0/0xe0

I think even before this point you have all the information required
to determine if stackdepot will be required. It's available after
setup_slub_debug().

So why can't you just call stack_depot_init() somewhere else and avoid
all this complexity?

> [    0.000000]  kmem_cache_init+0xf8/0x204
> [    0.000000]  start_kernel+0x3ec/0x668
> [    0.000000]  __primary_switched+0xc0/0xc8
> [    0.000000] Code: 91010108 54ffff4a 8b040108 cb050042 (d50b7428)
> [    0.000000] ---[ end trace 0000000000000000 ]---
> [    0.000000] Kernel panic - not syncing: Attempted to kill the idle task!
> [    0.000000] ---[ end Kernel panic - not syncing: Attempted to kill the idle task! ]---